Sunday, November 22, 2026 Edition No. 246
Search
Daily News Fully automated · 4 min read

AI News: November 22, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Anthropic Mathematician Levent Alpöge Says He Proved the Riemann Hypothesis With Claude, and Stanford Posts a Competing Human Draft Hours Later

Anthropic. Levent Alpöge, a mathematician employed by Anthropic, said on Friday that working with Claude he found a proof of the Riemann Hypothesis that also covers the Generalized Riemann Hypothesis. He posted the paper and a Lean formalization, split into 14 packages, on his own website rather than through an Anthropic announcement, and says the argument would run over 250,000 pages if printed. Scientific American reports the result was later verified by independent researchers, while The Decoder notes the remaining risk is that the problem statement was mistranslated into Lean. Hours after the announcement, AMS president Ravi Vakil posted an all-human draft proof of the GRH by Stanford’s Alessandro Masullo and later added that problems had been reported in that draft, found with the help of AI models; mathematicians are still checking both claims. Source

Also covered in today’s Anthropic roundup.

2. US Tech Stocks Post Their Worst Week Since the Pandemic as Nvidia Falls 14% and Anthropic Drops 9.1% a Day After Its Debut

Markets. The Nasdaq 100 fell 5.6% on Friday and 10.4% for the week, leaving it about 20% below its October peak and in bear market territory, according to the Financial Times as relayed by The Decoder. Nvidia dropped 14% after reports that the SEC is reviewing its accounting, the US semiconductor index fell 10.7%, and Anthropic fell 9.1% a day after its market debut. Doubts over AI buildout spending, the Nvidia situation, and macro pressures from interest rates and politics are cited as drivers. TechCrunch’s Equity podcast adds that Nvidia’s interim CEO Colette Kress said the company will stop selling H200 chips in China, while Elon Musk and others push for former CEO Jensen Huang’s return. Source

3. NVIDIA Researchers Argue Models Often Do Not Need Huge Context Windows, Since Periodic Summaries and Keyword Search Match Full Context

NVIDIA researchers argue that language models often do not need their full context windows, which are slow, expensive and prone to errors. In their tests, periodic summaries and targeted keyword search usually matched full-context results at a fraction of the cost, and the team proposes a new class of models that works from short notes instead of long context. The Decoder’s report is partly paywalled, so only the summary above was confirmed. Source

Also covered in today’s NVIDIA roundup.

4. ByteDance and US Universities Release Axion 1.0, an Open Agent Model Trained in Three Phases That They Say Rivals Top Commercial Models

ByteDance and several US universities introduced Axion 1.0, an AI agent model that the authors say rivals top commercial models from OpenAI and Anthropic on benchmarks after just three training phases. The paper argues that agent training pipelines have become so nested that it is hard to tell which step helps and expensive to reproduce, and it credits carefully curated training data rather than ever more training steps. The model, data and code are freely available. The benchmark comparison is the authors’ own. Source

5. OpenAI Publishes Prompt Templates for GPT-6.2 Astra and Sol and Warns That Editing Conversation History Now Drops Reasoning Items

OpenAI added copy-ready prompt templates to the GPT-6.2 Astra and Sol documentation to correct behaviors it observed, including overly brief answers, unnecessary follow-up questions and excessive testing when coding. On the API side, the Responses API drops reasoning items after prefix edits to conversation history, certain model changes including downgrades, and changes of end user or organization, which can lower answer quality and raise token use. Reusable instructions that should stay eligible for prompt caching belong in a developer message as the first input item, because top-level instructions are now appended after input. Source

Also covered in today’s OpenAI roundup.